hdminer - Bitcoin GPU miner in ATI CAL IL

History:
- 21 Dec 2010: first fully working version is completed, I start mining with it
- 28 Dec 2010: I advertise hdminer on IRC and sell the first copy (with source
  code, but license prohibits redistribution)
- 24 Jan 2011: I advertise hdminer more broadly on the forums at
  https://bitcointalk.org/?topic=2949
- mid-2011: hdminer becomes obsolete as the performance of open-source cgminer
  catches up with my implementation
- 04 Apr 2017: I release and place the code in the public domain, mostly as
  a historical artifact

Author: Marc Bevand, m.bevand at gmail.com
License: hdminer is hereby placed in the public domain

For installation instructions, see file "INSTALL".

Performance modelling.
  Blending:
    48 steps doing 13 ops each:
      5 ops to compute s0
      5 ops to compute s1
      3 adds to compute a word
  Steps:
    64 steps of 20 ops each:
      5 ops to compute s0
      2 ops to compute maj (optimized with BFI_INT to replace 5 ops)
      1 op to compute t2
      5 ops to compute s1
      1 op to compute ch (optimized with BFI_INT to replace 4 ops)
      4 ops to compute t1
      2 adds to update intermediate hash values
    8 adds to compute intermediate values
  Minus:
    3*20 because the last 3 steps of the second hash are not executed
    7 adds are not necessary in second hash because only H needs to be computed
Expected and measured bitcoin hash/sec:
  HD 6990@sw1 2703e9/((48*13+64*20+8)*2-3*20-7)/1e6 = 720M, measured 746M
  HD 6990@sw2 2550e9/((48*13+64*20+8)*2-3*20-7)/1e6 = 679M, measured 708M
  HD 5970     2320e9/((48*13+64*20+8)*2-3*20-7)/1e6 = 618M, measured 569M
  HD 6970     1352e9                                = 360M, est.     375M
  HD 5870     1360e9                                = 362M, measured 333M
  HD 6950     1127e9                                = 300M, measured 315M
  HD 5850     1044e9                                = 278M, est.     256M

Ideas:
- pick a merkle root so that its last word + s0[timestamp] = 0 to allow
  precomputing most of the blending
- precompute as much of the blending as possible for 1st SHA-256
- most of the blending for 2nd SHA-256 can be precomputed (zero words)
- getwork at least every 6sec 

data:
00000001 [---                         previous hash                 ---
--- ---] [---                         merkle root                   ---
--- ---] [tstamp] [tgtbit] [nonce-] 00000080 00000000 00000000 00000000
00000000 00000000 00000000 00000000 00000000 00000000 00000000 80020000